Add 64-bit Silvermont-optimized string/memory functions.

Add following functions:
bcopy, bzero, memcpy, memmove, memset, stpcpy, stpncpy, strcat, strcpy,
strlen, strncat, strncpy, memcmp, strcmp, strncmp.
Set all these functions as the default ones.

Change-Id: Ic66b250ad8c349a43d25e2d4dea075604f6df6ac
Signed-off-by: Varvara Rainchik <varvara.rainchik@intel.com>

1 files changed, 635 insertions, 0 deletions

diff --git a/libc/arch-x86_64/string/sse2-memmove-slm.S b/libc/arch-x86_64/string/sse2-memmove-slm.S
new file mode 100644
index 0000000..ee8440e
--- /dev/null
+++ b/libc/arch-x86_64/string/sse2-memmove-slm.S
@@ -0,0 +1,635 @@
+/*
+Copyright (c) 2014, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#include "cache.h"
+
+#ifndef MEMMOVE
+# define MEMMOVE		memmove
+#endif
+
+#ifndef L
+# define L(label)	.L##label
+#endif
+
+#ifndef cfi_startproc
+# define cfi_startproc	.cfi_startproc
+#endif
+
+#ifndef cfi_endproc
+# define cfi_endproc	.cfi_endproc
+#endif
+
+#ifndef cfi_rel_offset
+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+#endif
+
+#ifndef cfi_restore
+# define cfi_restore(reg)	.cfi_restore reg
+#endif
+
+#ifndef cfi_adjust_cfa_offset
+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#endif
+
+#ifndef ENTRY
+# define ENTRY(name)		\
+	.type name,  @function;		\
+	.globl name;		\
+	.p2align 4;		\
+name:		\
+	cfi_startproc
+#endif
+
+#ifndef END
+# define END(name)		\
+	cfi_endproc;		\
+	.size name, .-name
+#endif
+
+#define CFI_PUSH(REG)		\
+	cfi_adjust_cfa_offset (4);		\
+	cfi_rel_offset (REG, 0)
+
+#define CFI_POP(REG)		\
+	cfi_adjust_cfa_offset (-4);		\
+	cfi_restore (REG)
+
+#define PUSH(REG)	push REG;
+#define POP(REG)	pop REG;
+
+#define ENTRANCE	PUSH (%rbx);
+#define RETURN_END	POP (%rbx); ret
+#define RETURN		RETURN_END;
+
+	.section .text.sse2,"ax",@progbits
+ENTRY (MEMMOVE)
+	ENTRANCE
+#ifdef USE_AS_BCOPY
+	xchg	%rsi, %rdi
+#endif
+	mov	%rdi, %rax
+
+/* Check whether we should copy backward or forward.  */
+	cmp	%rsi, %rdi
+	je	L(mm_return)
+	ja	L(mm_len_0_or_more_backward)
+
+/* Now do checks for lengths. We do [0..16], [0..32], [0..64], [0..128]
+	separately.  */
+	cmp	$16, %rdx
+	jbe	L(mm_len_0_16_bytes_forward)
+
+	cmp	$32, %rdx
+	jg	L(mm_len_32_or_more_forward)
+
+/* Copy [0..32] and return.  */
+	movdqu	(%rsi), %xmm0
+	movdqu	-16(%rsi, %rdx), %xmm1
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, -16(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_len_32_or_more_forward):
+	cmp	$64, %rdx
+	jg	L(mm_len_64_or_more_forward)
+
+/* Copy [0..64] and return.  */
+	movdqu	(%rsi), %xmm0
+	movdqu	16(%rsi), %xmm1
+	movdqu	-16(%rsi, %rdx), %xmm2
+	movdqu	-32(%rsi, %rdx), %xmm3
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, 16(%rdi)
+	movdqu	%xmm2, -16(%rdi, %rdx)
+	movdqu	%xmm3, -32(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_len_64_or_more_forward):
+	cmp	$128, %rdx
+	jg	L(mm_len_128_or_more_forward)
+
+/* Copy [0..128] and return.  */
+	movdqu	(%rsi), %xmm0
+	movdqu	16(%rsi), %xmm1
+	movdqu	32(%rsi), %xmm2
+	movdqu	48(%rsi), %xmm3
+	movdqu	-64(%rsi, %rdx), %xmm4
+	movdqu	-48(%rsi, %rdx), %xmm5
+	movdqu	-32(%rsi, %rdx), %xmm6
+	movdqu	-16(%rsi, %rdx), %xmm7
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, 16(%rdi)
+	movdqu	%xmm2, 32(%rdi)
+	movdqu	%xmm3, 48(%rdi)
+	movdqu	%xmm4, -64(%rdi, %rdx)
+	movdqu	%xmm5, -48(%rdi, %rdx)
+	movdqu	%xmm6, -32(%rdi, %rdx)
+	movdqu	%xmm7, -16(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_len_128_or_more_forward):
+
+	cmp	$SHARED_CACHE_SIZE_HALF, %rdx
+	jae	L(mm_large_page_forward)
+
+	mov	%rsi, %r8  // copy src to r8
+	mov	%rdi, %r9  // copy dst to r9
+
+/* Aligning the address of destination.  */
+/*  save first unaligned 64 bytes */
+	movdqu	(%rsi), %xmm0
+	movdqu	16(%rsi), %xmm1
+	movdqu	32(%rsi), %xmm2
+	movdqu	48(%rsi), %xmm3
+
+	lea	64(%r9), %rdi
+	and	$-64, %rdi  /* rdi now aligned to next 64 byte boundary */
+
+	sub	%r9, %rsi /* rsi = src - dst = diff */
+
+	movdqu	(%rdi, %rsi), %xmm4
+	movdqu	16(%rdi, %rsi), %xmm5
+	movdqu	32(%rdi, %rsi), %xmm6
+	movdqu	48(%rdi, %rsi), %xmm7
+
+	movdqu	%xmm0, (%r9)
+	movdqu	%xmm1, 16(%r9)
+	movdqu	%xmm2, 32(%r9)
+	movdqu	%xmm3, 48(%r9)
+	movdqa	%xmm4, (%rdi)
+	movdqa	%xmm5, 16(%rdi)
+	movdqa	%xmm6, 32(%rdi)
+	movdqa	%xmm7, 48(%rdi)
+	add	$64, %rdi
+
+	lea	(%r9, %rdx), %rbx
+	and	$-64, %rbx
+
+	cmp	%rdi, %rbx
+	jbe	L(mm_copy_remaining_forward)
+
+	.p2align 4
+L(mm_main_loop_forward):
+
+	prefetcht0 128(%rdi, %rsi)
+
+	movdqu	(%rdi, %rsi), %xmm0
+	movdqu	16(%rdi, %rsi), %xmm1
+	movdqu	32(%rdi, %rsi), %xmm2
+	movdqu	48(%rdi, %rsi), %xmm3
+	movdqa	%xmm0, (%rdi)
+	movdqa	%xmm1, 16(%rdi)
+	movdqa	%xmm2, 32(%rdi)
+	movdqa	%xmm3, 48(%rdi)
+	lea	64(%rdi), %rdi
+	cmp	%rdi, %rbx
+	ja	L(mm_main_loop_forward)
+
+L(mm_copy_remaining_forward):
+	add	%r9, %rdx
+	sub	%rdi, %rdx
+/* We copied all up till %rdi position in the dst.
+	In %rdx now is how many bytes are left to copy.
+	Now we need to advance %r8. */
+	lea	(%rdi, %rsi), %r8
+
+L(mm_remaining_0_64_bytes_forward):
+	cmp	$32, %rdx
+	ja	L(mm_remaining_33_64_bytes_forward)
+	cmp	$16, %rdx
+	ja	L(mm_remaining_17_32_bytes_forward)
+	test	%rdx, %rdx
+	.p2align 4,,2
+	je	L(mm_return)
+
+	cmpb	$8, %dl
+	ja	L(mm_remaining_9_16_bytes_forward)
+	cmpb	$4, %dl
+	.p2align 4,,5
+	ja	L(mm_remaining_5_8_bytes_forward)
+	cmpb	$2, %dl
+	.p2align 4,,1
+	ja	L(mm_remaining_3_4_bytes_forward)
+	movzbl	-1(%r8,%rdx), %esi
+	movzbl	(%r8), %ebx
+	movb	%sil, -1(%rdi,%rdx)
+	movb	%bl, (%rdi)
+	jmp	L(mm_return)
+
+L(mm_remaining_33_64_bytes_forward):
+	movdqu	(%r8), %xmm0
+	movdqu	16(%r8), %xmm1
+	movdqu	-32(%r8, %rdx), %xmm2
+	movdqu	-16(%r8, %rdx), %xmm3
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, 16(%rdi)
+	movdqu	%xmm2, -32(%rdi, %rdx)
+	movdqu	%xmm3, -16(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_remaining_17_32_bytes_forward):
+	movdqu	(%r8), %xmm0
+	movdqu	-16(%r8, %rdx), %xmm1
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, -16(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_remaining_3_4_bytes_forward):
+	movzwl	-2(%r8,%rdx), %esi
+	movzwl	(%r8), %ebx
+	movw	%si, -2(%rdi,%rdx)
+	movw	%bx, (%rdi)
+	jmp	L(mm_return)
+
+L(mm_remaining_5_8_bytes_forward):
+	movl	(%r8), %esi
+	movl	-4(%r8,%rdx), %ebx
+	movl	%esi, (%rdi)
+	movl	%ebx, -4(%rdi,%rdx)
+	jmp	L(mm_return)
+
+L(mm_remaining_9_16_bytes_forward):
+	mov	(%r8), %rsi
+	mov	-8(%r8, %rdx), %rbx
+	mov	%rsi, (%rdi)
+	mov	%rbx, -8(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_len_0_16_bytes_forward):
+	testb	$24, %dl
+	jne	L(mm_len_9_16_bytes_forward)
+	testb	$4, %dl
+	.p2align 4,,5
+	jne	L(mm_len_5_8_bytes_forward)
+	test	%rdx, %rdx
+	.p2align 4,,2
+	je	L(mm_return)
+	testb	$2, %dl
+	.p2align 4,,1
+	jne	L(mm_len_2_4_bytes_forward)
+	movzbl	-1(%rsi,%rdx), %ebx
+	movzbl	(%rsi), %esi
+	movb	%bl, -1(%rdi,%rdx)
+	movb	%sil, (%rdi)
+	jmp	L(mm_return)
+
+L(mm_len_2_4_bytes_forward):
+	movzwl	-2(%rsi,%rdx), %ebx
+	movzwl	(%rsi), %esi
+	movw	%bx, -2(%rdi,%rdx)
+	movw	%si, (%rdi)
+	jmp	L(mm_return)
+
+L(mm_len_5_8_bytes_forward):
+	movl	(%rsi), %ebx
+	movl	-4(%rsi,%rdx), %esi
+	movl	%ebx, (%rdi)
+	movl	%esi, -4(%rdi,%rdx)
+	jmp	L(mm_return)
+
+L(mm_len_9_16_bytes_forward):
+	mov	(%rsi), %rbx
+	mov	-8(%rsi, %rdx), %rsi
+	mov	%rbx, (%rdi)
+	mov	%rsi, -8(%rdi, %rdx)
+	jmp	L(mm_return)
+
+/* The code for copying backwards.  */
+L(mm_len_0_or_more_backward):
+
+/* Now do checks for lengths. We do [0..16], [0..32], [0..64], [0..128]
+	separately.  */
+	cmp	$16, %rdx
+	jbe	L(mm_len_0_16_bytes_backward)
+
+	cmp	$32, %rdx
+	jg	L(mm_len_32_or_more_backward)
+
+/* Copy [0..32] and return.  */
+	movdqu	(%rsi), %xmm0
+	movdqu	-16(%rsi, %rdx), %xmm1
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, -16(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_len_32_or_more_backward):
+	cmp	$64, %rdx
+	jg	L(mm_len_64_or_more_backward)
+
+/* Copy [0..64] and return.  */
+	movdqu	(%rsi), %xmm0
+	movdqu	16(%rsi), %xmm1
+	movdqu	-16(%rsi, %rdx), %xmm2
+	movdqu	-32(%rsi, %rdx), %xmm3
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, 16(%rdi)
+	movdqu	%xmm2, -16(%rdi, %rdx)
+	movdqu	%xmm3, -32(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_len_64_or_more_backward):
+	cmp	$128, %rdx
+	jg	L(mm_len_128_or_more_backward)
+
+/* Copy [0..128] and return.  */
+	movdqu	(%rsi), %xmm0
+	movdqu	16(%rsi), %xmm1
+	movdqu	32(%rsi), %xmm2
+	movdqu	48(%rsi), %xmm3
+	movdqu	-64(%rsi, %rdx), %xmm4
+	movdqu	-48(%rsi, %rdx), %xmm5
+	movdqu	-32(%rsi, %rdx), %xmm6
+	movdqu	-16(%rsi, %rdx), %xmm7
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, 16(%rdi)
+	movdqu	%xmm2, 32(%rdi)
+	movdqu	%xmm3, 48(%rdi)
+	movdqu	%xmm4, -64(%rdi, %rdx)
+	movdqu	%xmm5, -48(%rdi, %rdx)
+	movdqu	%xmm6, -32(%rdi, %rdx)
+	movdqu	%xmm7, -16(%rdi, %rdx)
+	jmp	L(mm_return)
+
+L(mm_len_128_or_more_backward):
+
+	cmp	$SHARED_CACHE_SIZE_HALF, %rdx
+	jae	L(mm_large_page_backward)
+
+/* Aligning the address of destination. We need to save
+	16 bits from the source in order not to overwrite them.  */
+	movdqu	-16(%rsi, %rdx), %xmm0
+	movdqu	-32(%rsi, %rdx), %xmm1
+	movdqu	-48(%rsi, %rdx), %xmm2
+	movdqu	-64(%rsi, %rdx), %xmm3
+
+	lea	(%rdi, %rdx), %r9
+	and	$-64, %r9 /* r9 = aligned dst */
+
+	mov	%rsi, %r8
+	sub	%rdi, %r8 /* r8 = src - dst, diff */
+
+	movdqu	-16(%r9, %r8), %xmm4
+	movdqu	-32(%r9, %r8), %xmm5
+	movdqu	-48(%r9, %r8), %xmm6
+	movdqu	-64(%r9, %r8), %xmm7
+
+	movdqu	%xmm0, -16(%rdi, %rdx)
+	movdqu	%xmm1, -32(%rdi, %rdx)
+	movdqu	%xmm2, -48(%rdi, %rdx)
+	movdqu	%xmm3, -64(%rdi, %rdx)
+	movdqa	%xmm4, -16(%r9)
+	movdqa	%xmm5, -32(%r9)
+	movdqa	%xmm6, -48(%r9)
+	movdqa	%xmm7, -64(%r9)
+	lea	-64(%r9), %r9
+
+	lea	64(%rdi), %rbx
+	and	$-64, %rbx
+
+/* Compute in %rdx how many bytes are left to copy after
+	the main loop stops.  */
+	mov 	%rbx, %rdx
+	sub 	%rdi, %rdx
+
+	cmp	%r9, %rbx
+	jb	L(mm_main_loop_backward)
+	jmp	L(mm_len_0_or_more_backward)
+
+	.p2align 4
+L(mm_main_loop_backward):
+
+	prefetcht0 -128(%r9, %r8)
+
+	movdqu	-64(%r9, %r8), %xmm0
+	movdqu	-48(%r9, %r8), %xmm1
+	movdqu	-32(%r9, %r8), %xmm2
+	movdqu	-16(%r9, %r8), %xmm3
+	movdqa	%xmm0, -64(%r9)
+	movdqa	%xmm1, -48(%r9)
+	movdqa	%xmm2, -32(%r9)
+	movdqa	%xmm3, -16(%r9)
+	lea	-64(%r9), %r9
+	cmp	%r9, %rbx
+	jb	L(mm_main_loop_backward)
+	jmp	L(mm_len_0_or_more_backward)
+
+/* Copy [0..16] and return.  */
+L(mm_len_0_16_bytes_backward):
+	testb	$24, %dl
+	jnz	L(mm_len_9_16_bytes_backward)
+	testb	$4, %dl
+	.p2align 4,,5
+	jnz	L(mm_len_5_8_bytes_backward)
+	test	%rdx, %rdx
+	.p2align 4,,2
+	je	L(mm_return)
+	testb	$2, %dl
+	.p2align 4,,1
+	jne	L(mm_len_3_4_bytes_backward)
+	movzbl	-1(%rsi,%rdx), %ebx
+	movzbl	(%rsi), %ecx
+	movb	%bl, -1(%rdi,%rdx)
+	movb	%cl, (%rdi)
+	jmp	L(mm_return)
+
+L(mm_len_3_4_bytes_backward):
+	movzwl	-2(%rsi,%rdx), %ebx
+	movzwl	(%rsi), %ecx
+	movw	%bx, -2(%rdi,%rdx)
+	movw	%cx, (%rdi)
+	jmp	L(mm_return)
+
+L(mm_len_9_16_bytes_backward):
+	movl	-4(%rsi,%rdx), %ebx
+	movl	-8(%rsi,%rdx), %ecx
+	movl	%ebx, -4(%rdi,%rdx)
+	movl	%ecx, -8(%rdi,%rdx)
+	sub	$8, %rdx
+	jmp	L(mm_len_0_16_bytes_backward)
+
+L(mm_len_5_8_bytes_backward):
+	movl	(%rsi), %ebx
+	movl	-4(%rsi,%rdx), %ecx
+	movl	%ebx, (%rdi)
+	movl	%ecx, -4(%rdi,%rdx)
+
+L(mm_return):
+	RETURN
+
+/* Big length copy forward part.  */
+
+L(mm_large_page_forward):
+/* Aligning the address of destination. We need to save
+	16 bits from the source in order not to overwrite them.  */
+
+	mov	%rsi, %r8
+	mov	%rdi, %r9
+
+	movdqu	(%rsi), %xmm0
+	movdqu	16(%rsi), %xmm1
+	movdqu	32(%rsi), %xmm2
+	movdqu	48(%rsi), %xmm3
+
+	lea	64(%r9), %rdi
+	and	$-64, %rdi      /* rdi = aligned dst */
+
+	sub	%r9, %rsi        /* rsi = diff */
+
+	movdqu	(%rdi, %rsi), %xmm4
+	movdqu	16(%rdi, %rsi), %xmm5
+	movdqu	32(%rdi, %rsi), %xmm6
+	movdqu	48(%rdi, %rsi), %xmm7
+
+	movdqu	%xmm0, (%r9)
+	movdqu	%xmm1, 16(%r9)
+	movdqu	%xmm2, 32(%r9)
+	movdqu	%xmm3, 48(%r9)
+	movntdq	%xmm4, (%rdi)
+	movntdq	%xmm5, 16(%rdi)
+	movntdq	%xmm6, 32(%rdi)
+	movntdq	%xmm7, 48(%rdi)
+	add	$64, %rdi
+
+	lea	(%r9, %rdx), %rbx
+	and	$-128, %rbx
+
+	cmp	%rdi, %rbx
+	jbe	L(mm_copy_remaining_forward)
+
+	.p2align 4
+L(mm_large_page_loop_forward):
+	movdqu	(%rdi, %rsi), %xmm0
+	movdqu	16(%rdi, %rsi), %xmm1
+	movdqu	32(%rdi, %rsi), %xmm2
+	movdqu	48(%rdi, %rsi), %xmm3
+	movdqu	64(%rdi, %rsi), %xmm4
+	movdqu	80(%rdi, %rsi), %xmm5
+	movdqu	96(%rdi, %rsi), %xmm6
+	movdqu	112(%rdi, %rsi), %xmm7
+	movntdq	%xmm0, (%rdi)
+	movntdq	%xmm1, 16(%rdi)
+	movntdq	%xmm2, 32(%rdi)
+	movntdq	%xmm3, 48(%rdi)
+	movntdq	%xmm4, 64(%rdi)
+	movntdq	%xmm5, 80(%rdi)
+	movntdq	%xmm6, 96(%rdi)
+	movntdq	%xmm7, 112(%rdi)
+	lea 	128(%rdi), %rdi
+	cmp	%rdi, %rbx
+	ja	L(mm_large_page_loop_forward)
+	sfence
+
+	add 	%r9, %rdx
+	sub 	%rdi, %rdx
+/* We copied all up till %rdi position in the dst.
+	In %rdx now is how many bytes are left to copy.
+	Now we need to advance %r8. */
+	lea 	(%rdi, %rsi), %r8
+
+	cmp	$64, %rdx
+	jb	L(mm_remaining_0_64_bytes_forward)
+
+	movdqu	(%r8), %xmm0
+	movdqu	16(%r8), %xmm1
+	movdqu	32(%r8), %xmm2
+	movdqu	48(%r8), %xmm3
+	movdqu	-64(%r8, %rdx), %xmm4
+	movdqu	-48(%r8, %rdx), %xmm5
+	movdqu	-32(%r8, %rdx), %xmm6
+	movdqu	-16(%r8, %rdx), %xmm7
+	movdqu	%xmm0, (%rdi)
+	movdqu	%xmm1, 16(%rdi)
+	movdqu	%xmm2, 32(%rdi)
+	movdqu	%xmm3, 48(%rdi)
+	movdqu	%xmm4, -64(%rdi, %rdx)
+	movdqu	%xmm5, -48(%rdi, %rdx)
+	movdqu	%xmm6, -32(%rdi, %rdx)
+	movdqu	%xmm7, -16(%rdi, %rdx)
+	jmp	L(mm_return)
+
+
+/* Big length copy backward part.  */
+L(mm_large_page_backward):
+/* Aligning the address of destination. We need to save
+	16 bits from the source in order not to overwrite them.  */
+
+	movdqu	-16(%rsi, %rdx), %xmm0
+	movdqu	-32(%rsi, %rdx), %xmm1
+	movdqu	-48(%rsi, %rdx), %xmm2
+	movdqu	-64(%rsi, %rdx), %xmm3
+
+	lea	(%rdi, %rdx), %r9
+	and	$-64, %r9
+
+	mov 	%rsi, %r8
+	sub 	%rdi, %r8
+
+	movdqu	-16(%r9, %r8), %xmm4
+	movdqu	-32(%r9, %r8), %xmm5
+	movdqu	-48(%r9, %r8), %xmm6
+	movdqu	-64(%r9, %r8), %xmm7
+
+	movdqu	%xmm0, -16(%rdi, %rdx)
+	movdqu	%xmm1, -32(%rdi, %rdx)
+	movdqu	%xmm2, -48(%rdi, %rdx)
+	movdqu	%xmm3, -64(%rdi, %rdx)
+	movntdq	%xmm4, -16(%r9)
+	movntdq	%xmm5, -32(%r9)
+	movntdq	%xmm6, -48(%r9)
+	movntdq	%xmm7, -64(%r9)
+	lea 	-64(%r9), %r9
+
+	lea 	128(%rdi), %rbx
+	and 	$-64, %rbx
+
+/* Compute in %rdx how many bytes are left to copy after
+	the main loop stops.  */
+	mov 	%rbx, %rdx
+	sub 	%rdi, %rdx
+
+	cmp	%r9, %rbx
+	jae	L(mm_len_0_or_more_backward)
+
+	.p2align 4
+L(mm_large_page_loop_backward):
+	movdqu	-64(%r9, %r8), %xmm0
+	movdqu	-48(%r9, %r8), %xmm1
+	movdqu	-32(%r9, %r8), %xmm2
+	movdqu	-16(%r9, %r8), %xmm3
+	movntdq	%xmm0, -64(%r9)
+	movntdq	%xmm1, -48(%r9)
+	movntdq	%xmm2, -32(%r9)
+	movntdq	%xmm3, -16(%r9)
+	lea 	-64(%r9), %r9
+	cmp	%r9, %rbx
+	jb	L(mm_large_page_loop_backward)
+	jmp	L(mm_len_0_or_more_backward)
+
+END (MEMMOVE)